# Context and Memory Become the Agent Bottleneck

As agent context windows grow past the size of the codebases they work on, the limiting factor keeps shifting from model capability toward specification quality and persistent memory, moving competitive advantage and margin toward the harness layer rather than the model.

- Conviction: 34 / 100 (weakening)
- Horizon: Emerging (watchlist)
- Tracking since: 2026-08-24T00:00:00.000Z
- Last updated: 2026-08-27T14:00:31.454Z
- Canonical: https://polylog.news/ai/trends/agent-context-engineering
- Publisher: Polylog
- Affected regions: Global

## Recent score history

- 2026-08-26: 36
- 2026-08-27: 34

## Recent evidence

- [confirms] Two Papers Argue the Constraint on Coding Agents Is Now Specification and Memory, Not Model Capability (2026-08-24): Two new papers converge on specification and memory, not model capability, as the binding constraint on coding agents: one proposes a spec-driven development lifecycle for million-token agents, the other pre-loads Claude Code instances with prior work instead of starting each session from an empty context — direct evidence the bottleneck has moved to the harness layer.